CN103810496A - 3D (three-dimensional) Gaussian space human behavior identifying method based on image depth information - Google Patents

Abstract

The invention discloses a 3D (three-dimensional) Gaussian space human behavior identifying method based on image depth information. The 3D Gaussian space human behavior identifying method based on the image depth information includes the steps: extracting human skeleton 3D coordinates in the depth information, normalizing the 3D coordinates and filtering joints with low human behavior identifying rate and redundant joints; building interest joint groups according to behaviors, checking human motion space characteristics based on Gaussian distance to perform AP (affinity propagation) clustering to obtain word lists of behavior characteristics, and performing data cleaning for the word lists; building human behavior conditional random field identifying models, and classifying the human behaviors according to the human behavior conditional random field identifying models. The 3D Gaussian space human behavior identifying method has high interference resistance for specific directions, skeleton sizes and space positions of a human body, high generalization capability for motion differences led in by different experiment individuals and excellent identifying capability for inhomogeneous similar behaviors.

Description

3D Gauss space human body behavior recognition methods based on picture depth information

Technical field:

The present invention relates to field of machine vision, particularly a kind of 3D Gauss space human body behavior recognition methods based on picture depth information.

Background technology:

Human body behavior in video is identified in the fields such as a lot of video monitorings, man-machine interaction, video recovery important application.Although within the past ten years, various countries experts and scholars have proposed a lot of methods, have obtained in the field a lot of breathtaking progress, and high-precision human body behavior identification is still a job that has challenge.One of reason be exactly human body behavior be a kind of dynamic actuation time of sequence, exercises boundary is fuzzy, even its action of same people also can be out of shape, even exercises are combined mutually, the while is carried out the generation of the middle situation that may be blocked in action.Human body cutting apart from background itself is exactly a difficult task, further aggravated the difficulty of behavior identification.

The depth camera of releasing in recent years provides millimetre-sized 3D depth information.This has reduced the difficulty that human body is cut apart to a great extent.For depth information, Shotton has proposed a kind of single pixel object identification method (Shotton based on Stochastic Decision-making forest classified device, J., et al.Real-Time Human Pose Recognition in Parts from Single Depth Images.in Computer Vision and Pattern Recognition (CVPR), 2011 IEEE Conference on.2011.), the method has been used for reference object identification theory, adopt the middle expression of a kind of human body more difficult motion estimation to be mapped as to the classification problem of simple pixel-oriented, and adopt the method for the local optimum based on average drifting to find the optimal estimation of each joint.Based on the method, can directly obtain human body 3D skeleton joint coordinate.Human action is a kind of hinge arrangement, given skeleton as shown in Figure 1, left figure is depth image, right figure is the corresponding skeletal graph picture that the single pixel object identification method based on Stochastic Decision-making forest classified device of Shotton proposition obtains, human visual system can easily judge its action, even if part joint is blocked.

But, with a lot of noises, even there is manifest error in the 3D joint based on monocular depth information estimator, especially in the situation that blocking, as both hands intersect, multiple human body is touched mutually etc.Based on this 3D joint reasoning, still can not guarantee human body behavior accuracy of identification.

Summary of the invention:

The present invention is in order to overcome defect in above-mentioned prior art, and a kind of 3D Gauss space human body behavior recognition methods based on picture depth information of robust is provided.

To achieve these goals, the invention provides following technical scheme:

3D Gauss space human body behavior recognition methods based on picture depth information, the steps include:

Step 1, for the depth information of every two field picture, adopts the single pixel object identification method based on Stochastic Decision-making forest classified device that Shotton proposes confirm human body and further obtain human body 3D joint coordinates;

Step 2, by described human body 3D joint coordinates data normalization;

Step 3, screening human synovial, filters low joint or the redundancy joint of human body behavior identification contribution;

Step 4, analyzes every class behavior, based on AP clustering algorithm, adds up in every class behavior, and the articulation point that joint space movement travel is outstanding, builds interest and close knot cluster;

Step 5, for every class behavior, closes knot cluster based on interest, calculates the 3D Gauss space characteristics of each action;

Step 6, adopts AP clustering algorithm, builds Gauss apart from core, the 3D Gauss space characteristics that projects to human action space is gathered for the classification of motion of n group, and obtain the cluster centre that represents every group of action;

Step 7, for every group of action, adopts the affiliated cluster centre of each action to build behavioural characteristic word list, and every group of action is carried out to data scrubbing preparation;

Step 8, builds human body behavior conditional random field models, and training sample, obtains human body behavior model of cognition;

Step 9, identifies new samples.

In technique scheme, in step 2, described human body 3D joint coordinates data normalization is comprised to skeleton limbs vector size normalizing, skeleton reference zero normalizing and skeleton direction normalizing.

The step of wherein said skeleton limbs vector size normalizing comprises:

A) selecting a human body 3D joint coordinates is master pattern;

B) keep each sample limb segment direction vector constant, each vector is zoomed to master pattern length;

C) take buttocks center as reference point, structure joint tree, moves each joint according to convergent-divergent length, and mobile vector is:

here Δ d _fifi ancestors' of present node mobile vector, ancestors' number that n is present node.

The step of wherein said skeleton reference zero normalizing comprises: be new coordinate reference space O ' at zero point, mobile skeleton take buttocks center.

The step of wherein said skeleton direction normalizing comprises:

A) select former coordinate system X-axis, make itself and left stern arrive the vector of right stern

parallel, take new coordinate reference space O ' at zero point as buttocks center construction straight line is perpendicular to new ground reference plane, obtain new coordinate reference space Z axis;

B) rotation skeleton, is mapped to new coordinate reference space by skeleton.

In technique scheme, step 3 retains the large joint set of human body behavior identification contribution by screening, and the joint set of reservation comprises 12 joints: head, left/right elbow, left/right wrist, left/right knee, left/right ankle, left/right stern, buttocks center.

In technique scheme, the step that step 4 builds interest pass knot cluster by AP algorithm is:

A) move distance in calculating each joint of consecutive frame, the coordinate that is located at certain joint in consecutive frame (i frame, i+1 frame) is respectively: (x _ik, y _ik, z _ik), (x _{i+1, k}, y _{i+1, k}, z _{i+1, k}), move distance d _ikfor:

d _ik ²＝(x _ik-x _i+1k) ²+(y _ik-y _i+1k) ²+(z _ik-z _i+1k) ²

B) cumulative all move distances obtain the total kilometres D in a joint _k:

$D_k=\underset{i=1}{\overset{n-1}{\mathrm{\Σ}}}{d_{\mathrm{ik}}}^2$

C) based on AP algorithm, specify cluster numbers, employing Euclidean distance is measuring similarity, calculates gained move distance all joints are divided into 3 classes according to previous step;

D) abandon the shortest joint of move distance, get two long classes of move distance as the higher joint of contribution degree, the interest that builds the behavior is closed knot cluster.

In technique scheme, the computation process of the 3D Gauss space characteristics of each action in step 5 is:

A) 3d space is divided into m × n × l (m, n, l ∈ Z) sub spaces, each joint must be in a sub spaces;

B) the subspace gaussian density in all the other 11 joints of calculating except buttocks center:

(1) to each joint, calculate its subspace gaussian density,

$p(X,u,\mathrm{\Σ})=\frac{1}{\sqrt{{\left(2\mathrm{\π}\right)}^n}\sqrt{\left|\mathrm{\Σ}\right|}}{e}^{-\frac{1}{2}(x-\mathrm{\μ}{)}^T{\mathrm{\Σ}}^{-1}(x-\mathrm{\μ})}$

Wherein X represents joint coordinates, and u represents center, subspace, and ∑ represents covariance matrix, makes ∑=d/3*n*I, and d is every sub spaces catercorner length here, and n is subspace number, and I is unit matrix;

(2) for normal distribution, 99% information is included in positive and negative 3 standard deviations (being d*n*I, n=3.5), and order is apart from subspace centre distance d _{joint, bin}> ε (ε=d; ) the corresponding subspace gaussian density p in joint (X, u, ∑)=0;

C) the subspace gaussian density in 11 of each action joints has formed sparse motion characteristic expression.In technique scheme, in step 6, Gauss apart from the building method of core is:

here x, y represents two stack features vectors, σ is standard deviation.

In step 6, the 3D Gauss space characteristics clustering method in human action space is:

A) adopt above-mentioned Gauss apart from core, calculate each group of action gaussian density characteristic similarity s (x, y);

B) for large numbers matrix, order similarity be 0, build sparse similar matrix;

C) get

for reference value, n is number of samples, according to sample, transmits by message, automatically determines cluster numbers, and the AP cluster of sparse matrix is supported in application, obtains the individual cluster centre action of k '.In technique scheme, in step 7, the building method of behavioural characteristic word list is:

A) replace all samples of former action sequence for center of a sample's action under it, obtain one group of vision word strings;

B) clear up each behavior sample vision word strings, delete the word repeating continuously, to reduce the impact that between different samples, time migration causes, obtain behavioural characteristic word list.

In technique scheme, the human body behavior model of cognition that step 8 is obtained adopts PSS to be optimized:

min _θf(θ)＝-logp _θ(Y|X)+r(θ)，

Here

$r\left(\mathrm{\θ}\right)={\mathrm{\λ}}_1\underset{i}{\mathrm{\Σ}}\left|{\mathrm{\θ}}_i\right|+{\mathrm{\λ}}_2\sqrt{{\mathrm{\θ}}^2}$

$p_{\mathrm{\θ}}\left(Y\right|X)=\frac{{\mathrm{\Π}}_{t=1}^{T+1}\exp ({\mathrm{\Ψ}}_{\mathrm{\θ}}(y_t,X)+{\mathrm{\Ψ}}_{\mathrm{\θ}}(y_t,y_{t-1}))}{Z_{\mathrm{\θ}}\left(X\right)}$

${\mathrm{\Ψ}}_{\mathrm{\θ}}(y_t,X)={\mathrm{\Σ}}_{a=1}^A{\mathrm{\λ}}_a{h}_a(y_t,X),$

${\mathrm{\Ψ}}_{\mathrm{\θ}}(y_t,y_{t-1})={\mathrm{\Σ}}_{b=1}^B{\mathrm{\β}}_b{g}_b(y_t,y_{t-1}),$

$Z_{\mathrm{\θ}}\left(X\right)={\mathrm{\Σ}}_Y{\mathrm{\Π}}_{c\∈C(Y,X)}{\mathrm{\φ}}_{\mathrm{\θ}}^c(Y_c,X_c),$

for an implicit function corresponding to (clique) C,

$h_a(y_t,X)=1[y_t=m]x_j^i,$ Wherein m ∈ Y, j ∈ [0, T],

G _b(y _t, y _t-1)=1[y _t=m ₁∧ y _t-1=m ₂], wherein m ₁, m ₂∈ Y

Compared with prior art, the present invention has following beneficial effect:

By described human body 3D joint coordinates normalization technology, strengthen direction unchangeability, bone size constancy, the locus anti-interference of the method by step 2; Select by the large selection of joint set and the interest joint mass selection relevant to behavior of step 4 of step 3 human body behavior identification contribution, enlarged markedly distance of all categories, the interference that effectively the irrelevant joint of filtering causes, has strengthened the anti-noise ability of model; The learning system that selection, interest joint mass selection in conjunction with human body 3D joint coordinates normalization technology, joint set that identification contribution is large selected, 3D Gauss space characteristics sparse expression and human body behavior conditional random field models have built a robust jointly; The present invention has very high identification to generic behavior, and the individual action difference of introducing of different experiments is had to very strong generalization ability, similar behavior is also had to good recognition capability simultaneously.

Accompanying drawing explanation:

Fig. 1 is the skeleton schematic diagram that adopts the single pixel object identification method based on Stochastic Decision-making forest classified device of Shotton proposition to obtain;

Fig. 2 is the process flow diagram of the 3D Gauss space human body behavior recognition methods based on picture depth information of the present invention;

Fig. 3 is human body 3D joint coordinates normalization schematic diagram of the present invention;

Fig. 4 is the human body behavior conditional random field models schematic diagram that the present invention adopts;

Fig. 5 is that the further refinement of human body behavior conditional random field models that the present invention adopts is explained, this figure waves as example with the right hand;

Fig. 6 is the identity confusion matrix of the present invention to 8 kinds of common behaviors.

Embodiment:

As shown in Figure 1, 2, step 1, for every frame depth information, adopt the single pixel object identification method based on Stochastic Decision-making forest classified device that Shotton proposes confirm human body and further obtain human body 3D joint coordinates, left figure is depth image, and right figure is the corresponding skeletal graph picture that said method obtains.

As shown in Figure 2,3, step 2, comprises skeleton limbs vector size normalizing, skeleton reference zero normalizing and skeleton direction normalizing by described human body 3D joint coordinates data normalization.

The step of wherein said skeleton limbs vector size normalizing comprises:

A) selecting a human body 3D joint coordinates is master pattern;

B) keep each sample limb segment direction vector constant, each vector is zoomed to master pattern length;

C) take buttocks center as reference point, structure joint tree, moves each joint according to convergent-divergent length, and mobile vector is:

here Δ d _fifi ancestors' of present node mobile vector, ancestors' number that n is present node.

The step of wherein said skeleton reference zero normalizing comprises: be new coordinate reference space O ' at zero point, mobile skeleton take buttocks center.

The step of wherein said skeleton direction normalizing comprises:

A) select former coordinate system X-axis, make itself and left stern arrive the vector of right stern parallel, take new coordinate reference space O ' at zero point as buttocks center construction straight line is perpendicular to new ground reference plane, obtain new coordinate reference space Z axis;

B) rotation skeleton, is mapped to new coordinate reference space by skeleton.

By by described human body 3D joint coordinates normalization technology, strengthen direction unchangeability, bone size constancy, the locus anti-interference of the method.

As shown in Figure 2, step 3, screening human synovial, filter low joint or the redundancy joint of human body behavior identification contribution, retain the large joint set of human body behavior identification contribution, the joint set of reservation comprises 12 joints: head, left/right elbow, left/right wrist, left/right knee, left/right ankle, left/right stern, buttocks center.

As shown in Figure 6, the every class behavior in step 4, step 5 refers to the typical behavior that set sequence forms.In experimental result of the present invention, comprise the recognition result of following 8 kinds of behaviors: height jettisonings and throw, frontly play, side is played, trot, tennis ball hitting, tennis service, golf ball-batting, pick up and throw.

As shown in Figure 2, step 4, analyzes every class behavior, based on AP clustering algorithm, adds up in each class behavior, and the articulation point that joint space movement travel is outstanding, builds interest and close knot cluster:

A) move distance in calculating each joint of consecutive frame, the coordinate that is located at certain joint in consecutive frame (i frame, i+1 frame) is respectively: (x _ik, y _ik, z _ik), (x _{i+1, k}, y _{i+1, k}, z _{i+1, k}), move distance d _ikfor:

d _ik ²＝(x _ik-x _i+1k) ²+(y _ik-y _i+1k) ²+(z _ik-z _i+1k) ²

B) cumulative all move distances obtain the total kilometres D in a joint _k:

$D_k=\underset{i=1}{\overset{n-1}{\mathrm{\Σ}}}{d_{\mathrm{ik}}}^2$

C) based on AP algorithm, specify cluster numbers, employing Euclidean distance is measuring similarity, calculates gained move distance all joints are divided into 3 classes according to previous step;

D) abandon the shortest joint of move distance, get two long classes of move distance as the higher joint of contribution degree, the interest that builds the behavior is closed knot cluster.

Select by the large selection of joint set and the interest joint mass selection relevant to behavior of step 4 of step 3 human body behavior identification contribution, enlarged markedly distance of all categories, the interference that effectively the irrelevant joint of filtering causes, has strengthened the anti-noise ability of model.

As shown in Figure 2, step 5, for every class behavior, closes knot cluster based on interest, calculates the 3D Gauss space characteristics of each action:

A) 3d space is divided into m × n × l (m, n, l ∈ Z) sub spaces, each joint must be in a sub spaces;

B) the subspace gaussian density in all the other 11 joints of calculating except buttocks center:

(1) to each joint, calculate its subspace gaussian density,

$p(X,u,\mathrm{\Σ})=\frac{1}{\sqrt{{\left(2\mathrm{\π}\right)}^n}\sqrt{\left|\mathrm{\Σ}\right|}}{e}^{-\frac{1}{2}(x-\mathrm{\μ}{)}^T{\mathrm{\Σ}}^{-1}(x-\mathrm{\μ})}$

Wherein X represents joint coordinates, and u represents center, subspace, and ∑ represents covariance matrix, makes ∑=d/3*n*I, and d is every sub spaces catercorner length here, and n is subspace number, and I is unit matrix.

(2) for normal distribution, 99% information is included in positive and negative 3 standard deviations (being d*n*I, n=3.5), and order is apart from subspace centre distance d _{joint, bin}> ε (ε=d; ) the corresponding subspace gaussian density p in joint (X, u, ∑)=0.

C) the subspace gaussian density in 11 of each action joints has formed sparse motion characteristic expression.As shown in Figure 2, step 6, adopts AP clustering algorithm, builds Gauss apart from core:

here x, y represents two stack features vectors, σ is standard deviation.

The 3D Gauss space characteristics that projects to human action space is gathered for the classification of motion of n group, and obtains the cluster centre that represents every group of action:

A) adopt above-mentioned Gauss apart from core, calculate each group of action gaussian density characteristic similarity s (x, y);

B) for large numbers matrix, order similarity be 0, build sparse similar matrix;

C) get

for reference value, n is number of samples, according to sample, transmits by message, automatically determines cluster numbers, and the AP cluster of sparse matrix is supported in application, obtains the individual cluster centre action of k '.

As shown in Figure 2, step 7, for every group of action, adopts the affiliated cluster centre of each action to build behavioural characteristic word list, and every group of action is carried out to data scrubbing preparation:

A) replace all samples of former action sequence for center of a sample's action under it, obtain one group of vision word strings;

B) clear up each behavior sample vision word strings, delete the word repeating continuously, to reduce the impact that between different samples, time migration causes, obtain behavioural characteristic word list.

As shown in Figure 2, step 8, builds human body behavior conditional random field models, and training sample, obtains human body behavior model of cognition;

Human body behavior conditional random field models as shown in Figure 4, y in Fig. 4 _tprediction discrete state, x _tfor random action variable, step 7 obtains visual signature word list.

Human body behavior conditional random field models explains as shown in Figure 5, and Fig. 5 is listed is exemplified as the right hand action of waving, and its interest is closed knot cluster and comprised right finesse, right hand elbow and head.

The human body behavior model of cognition that step 8 is obtained adopts PSS (Schmidt, M., Graphical ModelStructure Learning with L1-Regularization, 2010, UNIVERSITY OF BRITISH COLUMBIA.) be optimized:

min _θf(θ)＝-logp _θ(Y|X)+r(θ)，

Here

$r\left(\mathrm{\θ}\right)={\mathrm{\λ}}_1\underset{i}{\mathrm{\Σ}}\left|{\mathrm{\θ}}_i\right|+{\mathrm{\λ}}_2\sqrt{{\mathrm{\θ}}^2}$

$p_{\mathrm{\θ}}\left(Y\right|X)=\frac{{\mathrm{\Π}}_{t=1}^{T+1}\exp ({\mathrm{\Ψ}}_{\mathrm{\θ}}(y_t,X)+{\mathrm{\Ψ}}_{\mathrm{\θ}}(y_t,y_{t-1}))}{Z_{\mathrm{\θ}}\left(X\right)}$

${\mathrm{\Ψ}}_{\mathrm{\θ}}(y_t,X)={\mathrm{\Σ}}_{a=1}^A{\mathrm{\λ}}_a{h}_a(y_t,X),$

${\mathrm{\Ψ}}_{\mathrm{\θ}}(y_t,y_{t-1})={\mathrm{\Σ}}_{b=1}^B{\mathrm{\β}}_b{g}_b(y_t,y_{t-1}),$

$Z_{\mathrm{\θ}}\left(X\right)={\mathrm{\Σ}}_Y{\mathrm{\Π}}_{c\∈C(Y,X)}{\mathrm{\φ}}_{\mathrm{\θ}}^c(Y_c,X_c),$ for an implicit function corresponding to (clique) C,

$h_a(y_t,X)=1[y_t=m]x_j^i,$ Wherein m ∈ Y, j ∈ [0, T],

G _b(y _t, y _t-1)=1[y _t=m ₁∧ y _t-1=m ₂], wherein m ₁, m ₂∈ Y

As shown in Figure 2,6, step 9, identifies new samples, and the present invention has very high identification to generic behavior, and the individual action difference of introducing of different experiments is had to very strong generalization ability, similar behavior is also had to good recognition capability simultaneously.

The learning system that selection, interest joint mass selection in conjunction with human body 3D joint coordinates normalization technology, joint set that identification contribution is large selected, 3D Gauss space characteristics sparse expression and human body behavior conditional random field models have built a robust jointly; The present invention has very high identification to generic behavior, and the individual action difference of introducing of different experiments is had to very strong generalization ability, similar behavior is also had to good recognition capability simultaneously.

The present invention has wide practical use in fields such as video monitoring, man-machine interaction, video frequency searchings.

Disclosed is above only several specific embodiment of the present invention, and still, the present invention is not limited thereto, and the changes that any person skilled in the art can think of all should fall into protection scope of the present invention.

Claims (9)

1. the 3D Gauss space human body behavior recognition methods based on picture depth information, is characterized in that, comprises the following steps:

Step 1, for the depth information of every two field picture, adopts the single pixel object identification method based on Stochastic Decision-making forest classified device that Shotton proposes confirm human body and further obtain human body 3D joint coordinates;

Step 2, by described human body 3D joint coordinates data normalization;

Step 3, screening human synovial, filters low joint or the redundancy joint of human body behavior identification contribution;

Step 4, analyzes every class behavior, based on AP clustering algorithm, adds up in each class behavior, and the articulation point that joint space movement travel is outstanding, builds interest and close knot cluster;

Step 5, for every class behavior, closes knot cluster based on interest, calculates the 3D Gauss space characteristics of each action;

Step 6, adopts AP clustering algorithm, builds Gauss apart from core, the 3D Gauss space characteristics that projects to human action space is gathered for the classification of motion of n group, and obtain the cluster centre that represents every group of action;

Step 7, for every group of action, adopts the affiliated cluster centre of each action to build behavioural characteristic word list, and every group of action is carried out to data scrubbing preparation;

Step 8, builds human body behavior conditional random field models, and training sample, obtains human body behavior model of cognition;

Step 9, identifies new samples.

2. the 3D Gauss space human body behavior recognition methods based on picture depth information according to claim 1, is characterized in that: in described step 2, described human body 3D joint coordinates data normalization is comprised to skeleton limbs vector size normalizing, skeleton reference zero normalizing and skeleton direction normalizing;

The step of wherein said skeleton limbs vector size normalizing comprises:

A) selecting a human body 3D joint coordinates is master pattern;

B) keep each sample limb segment direction vector constant, each vector is zoomed to master pattern length;

C) take buttocks center as reference point, structure joint tree, moves each joint according to convergent-divergent length, and mobile vector is: here Δ d _fifi ancestors' of present node mobile vector, ancestors' number that n is present node;

The step of wherein said skeleton reference zero normalizing comprises: be new coordinate reference space O ' at zero point, mobile skeleton take buttocks center;

The step of wherein said skeleton direction normalizing comprises:

A) select former coordinate system X-axis, make itself and left stern arrive the vector of right stern

parallel, take new coordinate reference space O ' at zero point as buttocks center construction straight line is perpendicular to new ground reference plane, obtain new coordinate reference space Z axis;

B) rotation skeleton, is mapped to new coordinate reference space by skeleton.

3. the 3D Gauss space human body behavior recognition methods based on picture depth information according to claim 1, it is characterized in that: described step 3 retains the large joint set of human body behavior identification contribution by screening, and the joint set of reservation comprises 12 joints: head, left/right elbow, left/right wrist, left/right knee, left/right ankle, left/right stern, buttocks center.

4. the 3D Gauss space human body behavior recognition methods based on picture depth information according to claim 1, is characterized in that: the step that described step 4 builds interest pass knot cluster by AP algorithm is:

A) move distance in calculating each joint of consecutive frame, the coordinate that is located at certain joint in consecutive frame (i frame, i+1 frame) is respectively: (x _ik, y _ik, z _ik), (x _{i+1, k}, y _{i+1, k}, z _{i+1, k}), move distance d _ikfor:

d _ik ²＝(x _ik-x _i+1k) ²+(y _ik-y _i+1k) ²+(z _ik-z _i+1k) ²

B) cumulative all move distances obtain the total kilometres D in a joint _k:

$D_k=\underset{i=1}{\overset{n-1}{\mathrm{\Σ}}}{d_{\mathrm{ik}}}^2$

C) based on AP algorithm, specify cluster numbers, employing Euclidean distance is measuring similarity, calculates gained move distance all joints are divided into 3 classes according to previous step;

D) abandon the shortest joint of move distance, get two long classes of move distance as the higher joint of contribution degree, the interest that builds the behavior is closed knot cluster.

5. the 3D Gauss space human body behavior recognition methods based on picture depth information according to claim 1, is characterized in that: the computation process of the 3D Gauss space characteristics of each action in described step 5 is:

A) 3d space is divided into m × n × l (m, n, l ∈ Z) sub spaces, each joint must be in a sub spaces;

B) the subspace gaussian density in all the other 11 joints of calculating except buttocks center:

(1) to each joint, calculate its subspace gaussian density,

$p(X,u,\mathrm{\Σ})=\frac{1}{\sqrt{{\left(2\mathrm{\π}\right)}^n}\sqrt{\left|\mathrm{\Σ}\right|}}{e}^{-\frac{1}{2}(x-\mathrm{\μ}{)}^T{\mathrm{\Σ}}^{-1}(x-\mathrm{\μ})}$

Wherein X represents joint coordinates, and u represents center, subspace, and ∑ represents covariance matrix, makes ∑=d/3*n*I, and d is every sub spaces catercorner length here, and n is subspace number, and I is unit matrix;

(2) for normal distribution, 99% information is included in positive and negative 3 standard deviations (being d*n*I, n=3.5), and order is apart from subspace centre distance d _{joint, bin}> ε (ε=d; ) the corresponding subspace gaussian density p in joint (X, u, ∑)=0;

C) the subspace gaussian density in 11 of each action joints has formed sparse motion characteristic expression.

6. the 3D Gauss space human body behavior recognition methods based on picture depth information according to claim 1, is characterized in that: in described step 6, Gauss apart from the building method of core is:

here x, y represents two stack features vectors, σ is standard deviation.

7. the 3D Gauss space human body behavior recognition methods based on picture depth information according to claim 1, is characterized in that: in described step 6, the 3D Gauss space characteristics clustering method in human action space is:

A) adopt above-mentioned Gauss apart from core, calculate each group of action gaussian density characteristic similarity s (x, y);

B) for large numbers matrix, order

similarity be 0, build sparse similar matrix;

C) get

for reference value, n is number of samples, according to sample, transmits by message, automatically determines cluster numbers, and the AP cluster of sparse matrix is supported in application, obtains the individual cluster centre action of k '.

8. the 3D Gauss space human body behavior recognition methods based on picture depth information according to claim 1, is characterized in that: in described step 7, the building method of behavioural characteristic word list is:

A) replace all samples of former action sequence for center of a sample's action under it, obtain one group of vision word strings;

B) clear up each behavior sample vision word strings, delete the word repeating continuously, to reduce the impact that between different samples, time migration causes, obtain behavioural characteristic word list.

9. the 3D Gauss space human body behavior recognition methods based on picture depth information according to claim 1, is characterized in that: the human body behavior model of cognition that described step 8 obtains adopts PSS to be optimized:

min _θf(θ)＝-logp _θ(Y|X)+r(θ)，

Here

$r\left(\mathrm{\θ}\right)={\mathrm{\λ}}_1\underset{i}{\mathrm{\Σ}}\left|{\mathrm{\θ}}_i\right|+{\mathrm{\λ}}_2\sqrt{{\mathrm{\θ}}^2},$

$p_{\mathrm{\θ}}\left(Y\right|X)=\frac{{\mathrm{\Π}}_{t=1}^{T+1}\exp ({\mathrm{\Ψ}}_{\mathrm{\θ}}(y_t,X)+{\mathrm{\Ψ}}_{\mathrm{\θ}}(y_t,y_{t-1}))}{Z_{\mathrm{\θ}}\left(X\right)},$

${\mathrm{\Ψ}}_{\mathrm{\θ}}(y_t,X)={\mathrm{\Σ}}_{a=1}^A{\mathrm{\λ}}_a{h}_a(y_t,X),$

${\mathrm{\Ψ}}_{\mathrm{\θ}}(y_t,y_{t-1})={\mathrm{\Σ}}_{b=1}^B{\mathrm{\β}}_b{g}_b(y_t,y_{t-1}),$

$Z_{\mathrm{\θ}}\left(X\right)={\mathrm{\Σ}}_Y{\mathrm{\Π}}_{c\∈C(Y,X)}{\mathrm{\φ}}_{\mathrm{\θ}}^c(Y_c,X_c),$

for an implicit function corresponding to (clique) C,

$h_a(y_t,X)=1[y_t=m]x_j^i,$ Wherein m ∈ Y, j ∈ [0, T],

G _b(y _t, y _t-1)=1[y _t=m ₁∧ y _t-1=m ₂], wherein m ₁, m ₂∈ Y.

Images